ksadov
ksadov is the personal project and publishing site of Konstantine Sadov, a machine-learning researcher and software developer. The site says it compiles his longform writing, mainly about his projects, which span audio-model interpretability, speech-driven avatar animation, LLM agents, and creative software.
At a Glance
- Machine-learning and interpretability researchers
- Speech/audio AI developers
- Open-source AI and creative-coding developers
- 3D avatar and virtual-character creators
- +1 more
AI Tools by ksadov
(1)
FREUD
Audio Model Interpretability Toolkit
Discussions
No discussions yet
Be the first to start a discussion about ksadov
Latest News
voice2motion repository updated with evaluation tooling and README instructions
manifold-llm-bot received a reconnect-on-close update; repository describes the configurable LLM trading bot
Published 'Why does Nanachi have rectangular pupils?' on the ksadov site
Published 'AI agents for prediction market trading,' describing the Manifold LLM-agent experiment
Products & Services
Feature Retrieval, Editing and Understanding for Developers: an MIT-licensed codebase for discovering and analyzing intermediate activations in audio models. It trains sparse autoencoders and provides an interactive GUI for inspecting Whisper activations and learned features; a hosted demo is available.
An MIT-licensed codebase and demo for training a network that converts streaming speech audio into predicted face and head movements represented as ARKit blendshapes. It includes dataset preprocessing, training, ONNX export, inference, and evaluation tooling.
A configurable LLM-agent trading bot for Manifold Markets. It supports multiple LLM backends, web search, customizable DSPy/OpenAI-compatible programs, live trading, auto-sell rules, historical backtesting, and evaluation/optimization scripts.
A co-authored open-source real-time any-to-one voice-conversion model designed to run on consumer CPUs, using generative-adversarial training and knowledge distillation.
Market Position
ksadov is an individual open-source researcher/developer rather than a commercial company. FREUD is positioned as a focused, audio-specific interpretability and sparse-autoencoder toolkit, complementary to broader interpretability/feature-exploration ecosystems such as Neuronpedia and model-interpretability work from Anthropic and OpenAI. voice2motion and LLVC similarly occupy experimental open-source niches in speech-to-animation and low-latency voice conversion rather than competing as packaged enterprise products.
Leadership
Founders
Konstantine Sadov
Machine-learning researcher at Resemble AI; previously an ML Engineer at Koe AI and a Formal Verification Engineer at Apple. He holds a degree in Computational Biology from Cornell University.
Executive Team
Konstantine Sadov
Founder of FREUD; Machine Learning Researcher
ML Researcher at Resemble AI; previously ML Engineer at Koe AI and Formal Verification Engineer at Apple; Computational Biology graduate of Cornell University.
Founding Story
The available sources identify ksadov as Sadov's personal site and project identity rather than a separately incorporated company. His FREUD work grew out of his participation in the 2024 Mozilla Builders cohort and an interest in making audio-model internals more understandable through mechanistic interpretability.
Target Markets
- Machine-learning and interpretability researchers
- Speech/audio AI developers
- Open-source AI and creative-coding developers
- 3D avatar and virtual-character creators
- Developers experimenting with LLM agents and forecasting
- Mechanistic interpretability research for speech and audio models
- Inspecting, editing, and understanding Whisper internal features
- Speech-driven 3D avatar and facial animation
- Low-latency voice conversion on consumer CPUs
- Experimenting with LLM forecasting and prediction-market agents
- Educational, experimental, and creative open-source software projects